Hybrid Meta-Heuristic Algorithms For Solving Network Design Problem

Discrete Optimization
Hybrid meta-heuristic algorithms for solving network
design problem
Hossain Poorzahedy *, Omid M. Rouhani
Institute for Transportation Studies and Research, Sharif University of Technology, P.O. Box 11365-9313, Tehran, Iran
Received 26 November 2005; accepted 21 July 2006
Available online 15 November 2006
Abstract
Network design problem has been, and is, an important problem in transportation. Following an earlier effort in design-
ing a meta-heuristic search technique by an ant system, this paper attempts to hybridize this concept with other meta-heu-
ristic concepts such as genetic algorithm, simulated annealing, and tabu search. Seven hybrids have been devised and tested
on the network of Sioux Falls. It has been observed that the hybrids are more effective to solve the network design problem
than the base ant system. Application of the hybrid containing all four concepts on a real network of a city with over 2
million population has also proved to be more effective than the base network, in the sense of finding better solutions
sooner.
2006 Elsevier B.V. All rights reserved.
Keywords: Transportation; Network design; Ant system; Hybrid algorithm
1. Introduction
Transportation network design problem (NDP)
is the problem of improving or expanding transpor-
tation networks under resource constraint(s). More
specifically, NDP is the problem of choosing among
a set of projects so as to optimize certain objective(s)
under resource constraint(s). Objectives are (usu-
ally) related to user benefits (or costs), and con-
straints are related to various resources which
bring about such benefits at the cost of the operator
of the network.
Traditionally, when origin–destination (O/D)
demand is fixed (inelastic), the benefits of the users
of the network is measured by the reduction of the
total travel time accrued by the choice of certain
subset of the given project set, as it is proportional
to most user transportation costs (such as monetary
travel cost) as well. Moreover, most resources
needed to materialize the decisions may be provided
by monetary budget, and hence a single budget
resource replaces all resource constraints to intro-
duce NDP as follows.
Let, N(V,A) be a network of concern with V as
the set of nodes and A the set of links. Let y be
the (row) vector of project decisions yij, where
yij = 1/0 represents the acceptance/rejection of pro-
ject link (i,j), respectively. Let, Ay be the set of pro-
ject links, and Ay1 the set of project links which are
accepted in y: Ay1 = {(i,j) 2 Ay : yij = 1}. Each pro-
ject link (i,j) has a (construction) cost Cij, and the
0377-2217/$ - see front matter 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2006.07.038
*
Corresponding author.
E-mail address: Porzahed@sharif.edu (H. Poorzahedy).
European Journal of Operational Research 182 (2007) 578–596
www.elsevier.com/locate/ejor

total construction cost is limited to the budget level
B. Let, also, xij be the flow in link (i,j), and x the
(row) vector of link flows. Moreover, let xks
p be the
flow in path p, p 2 Pks, from origin k to destina-
tion s, (k,s) 2 P, where P is the set of origin–destina-
tion pairs, and Pks the set of paths from origin k to
destination s. Such flows are generated by O/D
demand for travel, dks
, from k to s, (k,s) 2 P.
Finally, let tij(xij) be the travel time function of link
(i,j), (i,j) 2 A [ Ay, representing average travel time,
which is assumed to be only a function of flow in
link (i,j), continuous, differentiable, Riemann inte-
grable, and strictly convex, defined for xij P 0. Then
the traditional NDP may be stated as follows:
ðNDPÞ Min
y
f ðyÞ ¼
X
ði;jÞ2A[Ay1
x
ijtijðx
ijÞ
s:t::
X
ði;jÞ2Ay
Cijyij 6 B; ð1Þ
yij ¼ 0=1;ði;jÞ 2 Ay; ð2Þ
x
is the user equilibrium ðUEÞ
flow in the network NðV ;A[Ay1Þ;
ð3Þ
where x* is the solution of the following problem.
For any given vector of y:
ðUEÞ Min
x
X
ði;jÞ2A[Ay1
Z xij

tijðuÞdu
s:t::
X
p2Pks
xks
p ¼ dks
; 8ðk; sÞ 2 P; ð4Þ
xks
p P 0; 8p 2 Pks; 8ðk; sÞ 2 P; ð5Þ
xij ¼
X
ðk;sÞ2P
X
p2Pks
xks
p dks
ij;p;
8ði; jÞ 2 A [ Ay1; ð6Þ
where dks
ij;p is 1 if link (i,j) 2 p, p 2 Pks, otherwise 0.
This is a well-known bi-level programming problem,
which is described well in the literature of transpor-
tation network design. Steenbrink (1974a), Wong
(1984), Magnanti and Wong (1984), and Yang and
Bell (1998) review a range of the network design
problem and solution algorithms.
NDP is a difficult problem to solve. This was
found soon after the early work of LeBlanc (1975)
who proposed a branch-and-bound procedure to
solve this non-convex programming problem. Thus,
a wide spectrum of approaches have been investi-
gated to assess the trade-off between accuracy and
speed of the solution. This is done by: (a) attaining
a convex programming problem through replacing
user equilibrium flow by system equilibrium flow
(Dantzig et al., 1979); (b) bringing the complexity
of the problem down to a linear function of user tra-
vel time by assuming constant link cost function
(Boyce et al., 1973; Holmberg and Hellstrand,
1998); (c) constraint relaxation, particularly those
regarding the integrality of the projects or integer
decision variables (Steenbrink, 1974b; Abdulaal
and LeBlanc, 1979; Dantzig et al., 1979); (d) decom-
position of the problem based on link type,
geographical area, or links of the network to over-
come the size of the problem (Steenbrink, 1974b;
Dantzig et al., 1979; Hoang, 1982; Solanki et al.,
1998), (e) expediting the solution procedure through
aggregation of the network by link and node
abstraction or extraction (Haghani and Daskin,
1983); (f) facilitating the design procedure using an
intrinsic approach which defines a new design prob-
lem that lacks the complexity of NDP (Yang and
Bell, 1998); (g) good approximate and heuristic pro-
cedures to solve the problem by, e.g., replacing the
objective function by another well-behaved function
(Poorzahedy and Turnquist, 1982), or other heuris-
tics to solve the problem (Chen and Sul Alfa,
1991); and (h) heuristic procedures of modern types,
such as Genetic algorithm (GA) (Yin, 2000), Simu-
lated Annealing (SA) (Friesz et al., 1992; Lee and
Yang, 1994), GA, SA, and Tabu Search (TS) (Can-
tarella et al., 2002), and Ant system (AS) (Poorzah-
edy and Abulghasemi, 2005). To observe brevity of
the discussion, the interested readers are referred to
the first three reviewing references and the last refer-
ence for more discussion and references in the area
of network design problem.
This paper, in the continuation of a previous
work, aims to present some hybrid algorithms
which enhance the performance of Ant System pre-
sented by Poorzahedy and Abulghasemi (2005) to
solve NDP. This paper is organized as follows:
First, relevant modern heuristic algorithms are
reviewed in Section 2. Next, the proposed hybrid
algorithms are presented and discussed in Section
3. Numerical examples to investigate the properties
of the proposed algorithms are given in Section 4.
Section 5 is devoted to present the results of the
application of the algorithms to a real case. Finally,
Section 6 concludes the paper.
2. Meta-heuristic algorithms
Modern heuristic techniques, also called meta-
heuristics are a family of procedures which benefit
H. Poorzahedy, O.M. Rouhani / European Journal of Operational Research 182 (2007) 578–596 579

some sort of intelligence in their search for finding the
solution of a problem. Although one may not prove
that the solutions found by most of them are optimal
solutions, experiences during the past two decades
have shown that they find the solution, or a very good
solution, rapidly and effectively. This is particularly
so, for the combinatorial problems, such as NDP.
Neural Network (NN), Simulated Annealing (SA),
Genetic algorithm (GA), Scatter Search (SS), Tabu
Search (TS), and Ant system (AS) are among these
modern heuristics (Onwubolu, 2002; Goldberg,
1989). Each of such techniques possesses certain spe-
cific intelligence to search for the solution of a prob-
lem. Clearly, a collection of such abilities enhance the
power of the search engine. This is why that attention
is being paid recently to hybrid algorithms which col-
lect some of these intelligent senses and procedures.
Cantarella et al. (2002) present one such endeavor
and Greistorfer (1998) introduces another.
It is not the intention of this paper to review
meta-heuristic algorithms. However, to make this
paper self-contained, a brief description of each
such algorithm which will be used in this paper will
be presented in what follows. For more detailed dis-
cussion, the interested readers are referred to the
appropriate references in the literature and those
cited in this paper.
2.1. Genetic algorithm
Genetic algorithm (GA) is a population-based
mechanism in the traditional procedure of which
every two parent solutions give birth to two child
(successor) solutions, transferring a new combina-
tion of genes in the form of new chromosomes to
them. Each chromosome is identified by genes
(decisions) accepting some values such as 1 (accep-
tance of a character), and 0 (rejection of that char-
acter), carrying a value which shows its fitness or
effectiveness (in solving the problem). The popula-
tion size is kept fixed, always made up of the best
valued individuals (chromosomes). Like the natural
phenomena in genetics, genes are mutated; and to
make sure that new generations never fall behind
their predecessors the chromosomes of the elites of
the population are always kept in the new genera-
tion. To have a workable genetic algorithm, one
needs to have (a) a mating procedure, (b) a chromo-
some value estimation procedure, and (c) supple-
mental procedures, such as mutation, and elite
preservence. A general version of this algorithm is
as follows:
Step 0. Initialize.
Step 1. Create a population of individuals.
Step 2. Compute the (fitness) values of the
individuals.
Step 3. If convergence criteria are satisfied, go to
step 7.
Step 4. Identify parent individuals randomly based
on a probability proportional to their fit-
ness values.
Step 5. Create the children of a pair of parent indi-
viduals by their crossovers.
Step 6. Choose chromosomes and genes for muta-
tion, and go to step 2.
Step 7. Stop.
There are several proposed procedures for imple-
menting each of the above steps. The readers are
referred to Goldberg (1989) for further readings in
this area.
To adapt this concept to network design problem,
one may define a gene to represent a project, taking a
value of 1 to show its presence in a network, and 0
otherwise. A chromosome is a vector of such genes,
whose fitness value is the amount of reduction of
the total travel time of the users in the network as a
result of the presence of the collection of the projects
in the chromosome (network). Pairs of chromosomes
break in one or more points, and appropriate seg-
ments of mating chromosomes interchanged to create
the respective off-springs. Population survival obey
the natural law of ‘‘Stronger creatures live longer’’.
2.2. Ant systems
Ants have remarkable capability in finding
sources of food through a cooperative process. Ants,
belonging to a colony, leave the nest and move ran-
domly around in the search of a nearby feeding
source. When such a source is discovered, they come
back to inform others about it, while depositing a
chemical substance, called pheromone, sensible by
ants, to mark the path leading to the food source
from the nest. Pheromone evaporates over time.
Hence, only trails with higher pheromone concentra-
tion persist in leading ants. This can happen when
other searching ants find the food and emphasize
this information by laying their pheromone on the
ground on the same trails. Thus, less frequently
visited (e.g., distant) food sources lose address by
evaporation of pheromone marker of the paths lead-
ing to them from the nest. This is the cooperative
mechanism that an ant colony converges towards
580 H. Poorzahedy, O.M. Rouhani / European Journal of Operational Research 182 (2007) 578–596

an optimal solution. Dorigo et al. (1991) and Dorigo
et al. (1996) describe this approach:
Let t be a time counter, s a tabu list index, i a
point (vertex) in a space (i = 1,. . .,n), sij(t) phero-
mone intensity on the edge joining i and j at time
t, and Dsij an increment of such intensity. Let, also,
m be the number of ants, tabuk(s) represent the ‘‘vis-
ited’’ vertex by ant k in its sth move, and q be the
pheromone evaporation rate, 6 q 6 1. A general
version of an ant system (AS) algorithm may be sta-
ted as follows (Dorigo et al., 1996):
Step 0. Initialization. t :¼ 0, sij(t) :¼ c and Dsij :¼ 0
for all (i,j). Place m ants on the n nodes.
Step 1. Move. s :¼ 1, and for k = 1 to m do: place the
starting vertex in tabuk, and repeat until ant
energy is exhausted ( a resource constraint
violation). Choose vertex j to move to with
probability pk
ij (as defined below). s :¼
s + 1. Insert j in tabuk (i.e., tabuk(s) = j).
Step 2. Pheromone update. For k = 1 to m do: move
ant k to its nest. Compute the value of the
objective function for ant k path, and the
respective pheromone increment Dsk
ij (as
defined below). Save the best solutions
found, and the respective characteristics.
For edge (i,j) and for k = 1 to m do:
Dsij :¼ Dsij þ Dsk
ij; ð7Þ
sijðt þ 1Þ :¼ q sijðtÞ þ Dsij; ð8Þ
t :¼ t þ 1; Dsij :¼ 0; for all ði; jÞ:
Step 3. Stopping criteria. If stopping criteria are
met (and stagnation behavior not
observed), stop. Else, empty tabu lists for
all ants, and go to step 1.
Let vij be a measure of the ‘‘visibility’’ of vertex j
from the vertex i, associated with the goodness of
visiting vertex j as seen from the vertex i (and
possibly, given the past history of the vertices visited
in tabuk(Æ)). One such measure may be defined as
the net benefit of visiting a vertex alone (irrespective
of visiting other vertices). Then, pk
ij is a choice
function (for ant k currently at vertex i, to move
to a vertex j) which utilizes visibility of j from i
(vij) and amount of pheromone on edge (i,j) (sij)
such as
pk
ij ¼
sijðtÞ
a
vb
ij=
P
m2F k
simðtÞ
a
vb
im; if j 2 F k;
0; otherwise;
(
ð9Þ
where Fk is the set of vertices not in the tabuk (i.e.,
the set of allowable vertices), and a and b are two
parameters representing the relative importance
of trail intensity (at time t) and the visibility
information.
To determine the amount of pheromone to be laid
on the edges, let Df(yk
) be the net benefit of the path
chosen by ant k in the space of concern, yk
. Then,
Dsk
ij :¼ Df ðyk
Þ if (i,j) 2 yk
, and 0 otherwise.
2.3. Simulated annealing
Simulated Annealing (SA) is an algorithm simu-
lating the annealing process of metals. It consists
of heating the metal to a suitable temperature,
‘‘soaking’’ it at that temperature for certain time
period, and then reducing its temperature at a spec-
ified rate, so as to create a specific molecular struc-
ture order for possessing certain mechanical
properties (Salmon et al., 2002).
Analogous to an annealing process, SA proce-
dure is based on a Monte-Carlo model to solve a
combinatorial optimization problem (Onwubolu,
2002). Any solution, yk
, to the combinatorial opti-
mization problem is a configuration of the problem
(matter) corresponding to a state of a solid material.
The objective function f(yk
) corresponds to energy
of a solid, and the control parameter c corresponds
to the temperature of the solid. SA is a procedure
which generates a sequence of configurations, each
being found by perturbing a current configuration
and accepted according to Metropolis criterion, to
reach an equilibrium configuration for a given tem-
perature (c); and then reduces the temperature until
a frozen configuration (with the lowest energy) is
achieved. This equilibrium configuration corre-
sponds to the solution of the problem. A general
statement of this algorithm follows (see, e.g.,
Onwubolu, 2002):
Step 0. Initialization. c is given a high value. Con-
sider an initial configuration i.
Step 1. Solution generation. Perturb the current
configuration, i, obtain configuration j, with
costs f(yi
) and f(yj
), and Dfij = f(yj
) f(yi
).
Step 2. Acceptance criterion. Accept yj
with the fol-
lowing probability (Metropolis criterion):
Pr(accepting yj
) = 1, if Dfij 6 0, and e
Dfij
c if
Dfij 0. (That is, yi
:¼ yj
if Dfij 6 0, or if a
generated random number in [0,1) is less
than e
Dfij
c .) Continue this process until the

new configuration reaches an equilibrium,
when the probability distribution of config-
uration yj
approaches Boltzman distribu-
tion given below:
Prðconfiguration ¼ yj
Þ ¼ qjðcÞ ¼
1
QðcÞ
e
Dfij
c ;
ð10Þ
where Q(c) is the normalizing constant.
Step 3. Change of temperature. Lower the value of
the parameter, c, in steps, while allowing
the system to reach equilibrium through
steps 1 to 2 for each temperature level.
Step 4. Stopping criterion. Terminate the process
when no improvement in the cost of the
alternative solutions exists (approximately).
Solution of the combinatorial optimization
is the final ‘‘frozen’’ configuration.
2.4. Hybrid algorithms
Various heuristic algorithms have different mech-
anisms as constituents of their intelligence. There
are two points worth emphasizing in this respect:
(a) Some of these algorithms may have better per-
formance to solve certain problems, because of the
structure of the problem which is more exposed to
the investigating algorithm. Moreover, one expects
(b) an algorithm with a union of such intelligent
mechanisms to perform better than the constituting
algorithms alone, and such hybrid algorithms
should be able to solve general problems more
efficiently.
This very belief has led many researchers to build
hybrid algorithms. Such endeavors include three
approaches: (a) One approach is to build parallel
heuristic techniques for solving combinatorial opti-
mization problems (Resende et al., 1999). (b) Cross-
ing two or more breeds of algorithms to create
hybrids for optimization is another approach. One
popular avenue of research in this respect is combin-
ing population-based procedures with local search
ones so as to gather two aspects of a good searching
technique, viz diversification and intensification,
together. This approach includes genetic algo-
rithm-tabu search hybrids for solving optimization
problems (e.g., Glover et al., 1995; Costa, 1995;
Greistorfer, 1998; Fleurent and Ferland, 1999);
genetic algorithm-simulated annealing hybrid solu-
tion procedures (e.g., Kim et al., 1994); simulated
annealing–case based reasoning (Sadek, 2001); and
tabu search and neural networks (De Werra and
Hertz, 1989). And, the third approach is (c) using
meta-heuristic solutions to start or guide other algo-
rithms (Chelouah and Siarry, 2003; Taillard and
Gambardella, 1997).
Hertz and Widmer (2003) present some guide-
lines for the use of meta-heuristics in solving combi-
natorial optimization problems, which may also be
used in creating such algorithms in hybrid forms.
3. The proposed hybrid algorithms
Following a previous endeavor by Poorzahedy
and Abulghasemi (2005), in presenting an AS for
solving transportation network design problems,
this paper tries to enhance the ability of such an
algorithm by hybrids which imbed additional intel-
ligence into the solution procedure. To do this, first
a base AS algorithm will be introduced, and next
some hybrid algorithms are presented, whose solu-
tion capabilities are measured based on their appli-
cations in solving one well-known test network of
Sioux Falls, and one large real case network.
3.1. The base AS algorithm
The base AS algorithm for solving NDP is intro-
duced in Poorzahedy and Abulghasemi (2005). A
graph is constructed as in Fig. 1, where the vertices
represent the projects (i), which are connected to
other projects (j) by undirected edges (i,j) with costs
equal to the construction cost of vertex j. Ants are
placed at starting vertices, called nests, to start mov-
ing from vertex to vertex in the hope of finding a
good basket of food, while their energy for search
last. They leave a pheromone trail on the edges,
and return to their nests, when energy-wise it is time
to return. This energy comes from the budget avail-
1 i
n n-1
3
2
Fig. 1. A graph representation of project selection for ants.

able to the ant, B, and will be spent by an amount cl,
when project l (including that of the nest) is picked.
Previous experiments show that, a good strategy is
having a number of ants equal to the number of
candidate projects, and let ant m possess nest m.
This would diversify the solutions (Poorzahedy
and Abulghasemi, 2005).
Ants find their paths from vertex (i) to vertex (j)
according to the visibility of vertex j from vertex i,
vij, and the concentration of pheromone on the edge
(i,j), sij. As in Poorzahedy and Abulghasemi (2005),
the choice of the next vertex, j, by an ant k currently
(at time t) at vertex i, follows a choice process gov-
erned by a multinomial logit:
pk
ijðtÞ ¼
euijðtÞ
P
m2F k
euimðtÞ
; if j 2 F k;
0; otherwise;
8

:
ð11Þ
where Fk is as defined in Eq. (9), and uij(t) is the util-
ity of choosing j from i at time t. At time t, the util-
ity of vertex j, among such vertices allowable to
choose from by ant k in Fk, is a function of the pher-
omone concentration on the edge leading i to j,
sij(t), and other variables such as the goodness of
project j alone,
vij ¼ vj ¼ f ðy0
Þ fj; ð12Þ
where fj is the value of the objective function when
only project j is constructed in the network, and
f(y0
) is this value for the do nothing alternative.
Then, one may define the utility of choosing vertex
j, for example, as follows:
uijðtÞ ¼ a sijðtÞ þ b vj; ð13Þ
where a, and b are the relative importance of the
variables in the utility function, and are given
parameters.
After completing the tour and returning to the
nest, ant-gathered information may be transferred
to the edges of the design graph by computing the
objective function of the selected set of projects
(set of visited vertices), yk
, and computing the net
benefit of this set found by ant k, as follows:
sk
ijðtÞ ¼
½f ðy0
Þ f ðyk
Þ cck
; if ði; jÞ 2 Ak
;
0; otherwise;
(
ð14Þ
where Ak
is the set of edges in the path traversed by
ant k, and ck
is the total cost of the projects selected
by ant k. c is a constant which converts monetary
units into travel time saved (in vehicle-hours). For
the m ants sent each time before updating the edge
pheromone, one may write:
Dsij ¼
X
m
k¼1
Dsm
ij: ð15Þ
Then, for the next iteration t + 1, or t + m depend-
ing on the definition, the pheromone concentration
of the edges may be updated as follows:
sijðt þ 1Þ ¼ qsijðtÞ þ Dsij; ð16Þ
where 1 q is the evaporation rate of pheromone
trail, 0 q 1.
The following is an adaptation of the ant-cycle
algorithm (Dorigo et al., 1996) to the NDP,
referred to as the base algorithm. The parameters
of the algorithm are set equal to the best values
obtained by Poorzahedy and Abulghasemi (2005),
as m = no. of projects, q = 0.5, b = 1.0, c = 1.0,
and vertex k = nest of ant k. One other parameter,
a, is chosen appropriately, and kept fixed for each
run.
Algorithm 1. The base Ant System network design
algorithm
Step 0. Initialization.
• Set t :¼ 0; solve UE flow problem for
y = y0
and compute the objective function
of the design problem f(y0
). f* :¼ f(y0
),
y* :¼ y0
. • Set sij :¼ 0, and solve UE flow
problem for each project j and compute
the respective design objective function fj,
j = 1,. . .,m. Set iteration counter c :¼ 0.
Step 1. Move. • Empty tabu lists for all ants. • For
k = 1 to m, do: Set ant k on project vertex k,
and place vertex k in the respective tabu list.
• Repeat the followings until tabu list
becomes budget infeasible: specify the
allowable projects Fk; compute the utility
of each project j 2 Fk according to Eq.
(13), and then the respective probability of
choice according to Eq. (11); choose the
next project vertex using a Monte-Carlo
simulation based on the latter values.
• Next, identify the set of chosen projects
for ant k, and solve the UE flow problem
for the selected projects of vector y = yk
,
and compute the respective design objective
function f(yk
). (Save yk
and f(yk
) to avoid
re-computation of same information in the
future.) If f(yk
) f*, set f* :¼ f(yk
) and
y* :¼ yk
.

Step 2. Pheromone Update. • For each ant k = 1 to
m do: Compute pheromone increment
according to Eq. (14). • Add pheromone
increments for all ants to compute the total
such increment Dsij as in Eq. (15), and
update the pheromone of each edge (i,j)
according to Eq. (16). • Set t :¼ t + 1, and
Dsij :¼ 0 for all (i,j).
Step 3. Stopping Criteria. Set c :¼ c + 1, and if
c 6 c0
(say 10) go to step 1, otherwise y* is
the best solution found so far, with design
objective function value f(y*). Collect the
necessary information, and stop.
A description of the algorithm may be found in
Poorzahedy and Abulghasemi (2005). However,
before passing it is worth noting that since the
presence of all projects in the path found by an ant k
in the search graph contribute to the total benefit, it
is advisable to lay pheromone on all edges of the
graph linking these projects (and not only the edges
in the path of ant k). As an example, if projects i, j,
and k are found by an ant in a path i–j–k, then
pheromone is laid on edges (i,j), (i,k), and (j,k), and
not only the edges (i,j) and (j,k) in the path. This is
the practice in this study.
3.2. Improvement mechanisms of the base algorithm
3.2.1. (I) Improvement 1
Dorigo and Gambardella (1997) present an Ant
Colony System (ACS) which follows a rule similar
to the following in choosing the vertex among the
allowable ones for ant k:
j ¼
arg maxm2F k
fuimg; if r 6 r0;
J; otherwise;
8

:
ð17Þ
where r is a random variable with uniform distribu-
tion in [0,1], and r0 is a pre-specified parameter ; and
J is a random variable selected according to Eq. (11)
(Also, see Maniezzo et al., 2003). Incorporating (17)
and (11) into ant’s decisions, instead of (11) alone,
favors projects with higher utilities (higher phero-
mone trails, higher visibility, lower cost, or higher
goodness or values).
The following improvement mechanism has been
devised to favor solutions which are stronger, with-
out reducing the width of the spectrum of the solu-
tions that could be found by the ants, but only
emphasizing the stronger solutions.
Imp.1: In step 2 of the base algorithm: • Find the
top 3 solutions found by the m ants (which have the
least values of the design objective function). Con-
sider pheromone updating only for ants k corre-
sponding to these 3 solutions (instead of the m
solutions of all ants).
3.2.2. (II) Improvement 2
Mutation may be viewed as a local search in the
solution space. In another experiment, a mutation
mechanism has been devised to perform such a
search in the vicinity of the preferred solutions, with
the hope that they have higher potential to reveal
better solutions than others. With such intention,
then, naturally these mutations must be motivated
consciously based on the available information,
rather than randomly. In order to make sure that
the available information is mature enough and
valid to guide such search, mutation is performed
in mid-iteration numbers (here iteration 5) of the
algorithm. Since the mutated solutions should be
feasible resource-wise, accepting a project which is
rejected in the solution under mutation operation,
should coincide with rejection of one (or more)
accepted project(s) depending on the resource use,
or costs, of the projects.
The second proposed improvement mechanism
borrows ideas from local search techniques such as
tabu search, and takes advantage of a measure of
relative merit for projects such as the frequency of
presence in the solutions found. The improvement,
which is a kind of embedding a local search mecha-
nism within a population search procedure, is as
follows:
Imp.2: In the 5th iteration of the base algorithm,
replace the usual operations in step 1 by the follow-
ing mutation operation. Find the set of top 3 solu-
tions of each of the previous iterations skip
repeated solution and choose the next best solution,
so that a maximum of 12 different solutions are
obtained. Compute the frequency of the presence
of the genes (projects) in the solution chromosomes
(vectors) in this set. Next, find the set of top 2 solu-
tions of each, and all, of previous iterations (maxi-
mum of m = 10 solutions, and repeated solutions
are allowed). For each (repeated) solution in this
set, replace the project with the (next) least fre-
quency of presence, with the (next) most frequently
present project not chosen in that solution, while
keeping the solution feasible, and making sure that
the new solution has not been found and examined
before. Call the new set of solutions the mutated set.

Perform the UE traffic assignment procedure for
the mutated set of solutions, Compute the respective
objective functions, and treat them as the solutions
found by (the m=) 10 ants in an iteration of the
algorithm. Save the solution information (yk
,f(yk
)),
and if f(yk
) f*, set f* :¼ f(yk
) and y* :¼ yk
.
Next, perform step 2 (pheromone updating) of
the algorithm, as usual.
Non-random mutation operations have been
exploited by the other researchers, as well. Cree
et al. (1998) exploits gradient of the objective func-
tion to guide local search. Fleurent and Ferland
(1999) have, also, combined genetic mutation con-
cept with ideas in tabu search. Such ideas of project
replacements are also present in earlier works in
project selection problems. Ahmad (1983) has
exploited such an idea in his gradient projection
technique in choosing road maintenance projects.
3.2.3. (III) Improvement 3
The two improvement mechanisms (1 and 2) dis-
cussed above favor strong genes, and species and
genes, respectively. The next improvement mecha-
nism borrows ideas from Simulated Annealing
(SA) to reduce the computational burden of the
search. Kim et al. (1994) have brought in the idea
of controlling the population of the weaker species
by exploiting Metropolis Criterion.
In solving NDP, the major computational bur-
den of the solution procedure is the time it takes
to solve a UE flow problem to compute the objec-
tive function value for a given solution yk
. To pre-
vent solving the latter problem for a seemingly (or
potentially) irrelevant solution, one may take
advantage of this idea that information regarding
the fitness of a solution increases as the procedure
proceeds. SA procedure provides such an increas-
ingly tighter sieves to sift coarser entities, while pro-
viding higher diversity at the beginning.
To implement this idea, define g1(n,k) as the
goodness of the combination of the projects in a
solution found by ant k, yk
, in iteration n, to be
equal to the sum of the pheromone laid upon the
edge joining each pair of the projects accepted in
yk
. The more each pair of projects reveal themselves
in solutions of the m ants in a given iteration, the
higher the pheromone laid on the edge joining them.
(This concept may be conceivably generalized for
each p-project combination, p = 2,3,. . .) And, hence
pairs of projects which accentuate the net benefit of
each other more than other combinations of pro-
jects happen to occur more frequently in the solu-
tions found by the ants, which in turn reflect
themselves into measure g1(.) through sij. Thus, let
g1ðn; kÞ ¼
X
m
s¼1
X
m
t¼sþ1
a sðtabukðsÞ; tabukðtÞÞ; ð18Þ
where tabuk(s) contains the index of the project
visited in the sth move of ant k, s = 1,2,. . . If
tabuk (s) = i and tabuk(t) = j, then sij is the phero-
mone on the edge joining them. a is a scaling con-
stant in Eq. (18). Let g1ðnÞ be the average value of
such goodness measures for all ants in iteration n:
g1ðnÞ ¼
1
m
X
m
k¼1
g1ðn; kÞ: ð19Þ
Define g1ðnÞ as the moving average of the two pre-
vious values of g1ð:Þ:
g1ðnÞ ¼ ½g1ðn 1Þ þ g1ðn 2Þ=2; ð20Þ
g1(n,k) in Eq. (18) may be envisaged as a measure of
the energy level of the configuration (of the network
or projects) found by ant k in iteration n, and g1ðnÞ
and g1ðnÞ as the respective average values defined
according to Eqs. (19) and (20). Define the change
in energy levels of the configuration found by ant
k in iteration n as compared to an average value
(here, of the two previous average energy levels),
g1ðnÞ, be as follows:
DEðn; kÞ ¼ g1ðn; kÞ g1ðnÞ: ð21Þ
Thus, analogous to the Metropolis Criterion in SA
algorithm, a solution yk
with lower energy level
(goodness) than the average g1ð:Þ will only be ac-
cepted with probability of e
DEðkÞ
C . (Note that, unlike
the reduction of energy level in simulated annealing,
we would like to see the energy of a configuration
found by an ant to increase in our case. Of course,
there is a limit to this energy level, which is defined
by the optimal configuration. If a solution is re-
jected we may save the computation time of solving
the respective UE flow problem. Thus, the proposed
improvement in the base algorithm may be stated as
follows.
Imp.3: In step 1 of the base algorithm, in iteration
n, n 2: Compute DE(k) = DE(n,k) in Eq. (21).
When the solution of ant k, yk
, is found, the UE
flow problem is solved for this solution only if
DE(k) 0, and if DE(k) 6 0 this problem is solved
with a probability of Pr{accepting yk
} = eDE(k)/c(n)
,
where c(n) is a decreasing function of n (e.g.,
c(n) = 0.9c(n 1)). (To check the acceptability of
yk
with non-positive DE, a uniformly distributed

random variable r in [0,1] is generated, and yk
is
accepted if r eDE(k)/c(n)
.)
An alternative measure of the goodness of the
configuration (solution) yk
is defined as follows.
Let, Nðn; yk
l Þ be the frequency of accepting project
l (=1 to the number of projects = m, the number
of ants, here) in the solutions found by the m ants
in iteration n. Let,
g2ðn; kÞ ¼
X
m
l¼1
Nðn; yk
l Þ ð22Þ
be the sum of such frequencies of the projects ac-
cepted in yk
. Defining, as before:
g2ðnÞ ¼
1
m
X
m
k¼1
g2ðn; kÞ; ð23Þ
g2ðnÞ ¼ ½g1ðn 1Þ þ g1ðn 2Þ=2; ð24Þ
DEðn; kÞ ¼ g2ðn; kÞ g2ðnÞ; ð25Þ
a similar procedure of Imp.3 may be run using this
new definition of DE(Æ).
3.2.4. (IV) Improvement 4–7
Each improvement in (I) through (III) above
brings a new variation of the base algorithm,
increasing its efficiency in solving NDP. It would
be interesting to see the resulting effect of hybrids
of such improvements, as follows:
Imp.4: Imp.2 + Imp.3, Imp.5: Imp.1 + Imp.2,
Imp.6: Imp.1 + Imp.3, Imp.7: Imp.1 + Imp.2 +
Imp.3.
4. Numerical example
To show the performance of the proposed algo-
rithms, they will be applied to the case of the net-
work of Sioux Falls. This network, shown in
Fig. 2, has 24 nodes and 76 links. There are 10 pairs
of project links, of which 5 projects are improve-
ments on the existing links, and 5 are new links.
The parameters of the average travel time functions
tij ¼ aij þ bijx4
ij, for link (i,j), and the origin/destina-
tion (O/D) demand, are basically those given in
Poorzahedy and Turnquist (1982) and LeBlanc
(1975), and are omitted here for brevity. The con-
struction costs of projects 1 to 10 are, respectively,
625, 650, 850, 1000, 1200, 1500, 1650, 1800, 1950,
and 2100 units of money. To measure the accuracy
of the solutions given by the algorithms, all possible
combinations of 210
(=1024) different networks
have been created for the network under study, in
order to identify the exact solution of the problem
under various budget levels. Since the meta-heuristic
algorithms of concern are of stochastic nature, each
case has been solved 50 times, and the average mea-
sures of such runs have been reported as the mea-
sure of performance of the algorithms. The
average CPU time for solving one UE flow problem
is under 0.1 seconds of a personal computer equiva-
lent to Pentium IV, 1400 MHz.
4.1. Application of base AS algorithm
First the performance of the base AS algorithm
will be discussed. In all experiments the following
parameters are set as shown, unless otherwise spec-
ified: m = jAyj = 10, q = 0.5, b = 1, c = 1, nest of
ant k = project k. Table 1 shows the results of solv-
ing the NDP for the test network under various
budget levels, as measured by budget to total project
cost ratio (B/C). This ratio shows the level of limita-
tion of the budget in the design problem, a determi-
nant of the level of efforts needed to solve the
problem, as shown in Table 1. In this table the cost
of solving the problem is measured by the CPU time
of the computer, which is (almost) directly propor-
tional to the number of UE flow problem (assign-
ments) solved. The performance of the algorithm
is measured by the frequency of finding the optimal
solution in 50 runs of the algorithms to solve the
same problem. The worst and best values of the
objective function of the design problem also show
the level of ‘‘cost’’ of non-optimal solution found
by the algorithm.
Table 1 shows that no. of assignments increases
as the B/C increases from a low value of 0.2 to a
mid value of 0.5, and then decreases when B/C
reaches a high value of 0.8, a phenomenon expected
to occur because of the level of feasible and domi-
nating alternative networks which has similar varia-
tion as the no. of assignments.
The following experiments have some interesting
results. Suppose projects are of equal cost (this is no
restriction), and that budget allows choosing 5 pro-
jects. Fig. 3 shows that increasing a from 0 to 0.5
reduces the average no. of assignments (in 50 runs)
required to solve the problem. This is because
increasing a increases the effect of pheromone con-
centration on the choice process in Eq. (11), which
means better and more frequently chosen projects
will be chosen more, and hence there will be less

diversity in the solution set found by various ants.
This results lower no. of assignments for higher
value of a. This observation suggests the use of
lower a values for lower B/C levels for which alter-
native feasible solutions become more diverse, and
examining them requires lower a values.
Two experiments have been designed to investi-
gate the effect of the goodness of project j, vj, upon
the efficiency of the algorithm. For the above 5 pro-
ject budget example with B/C = 0.5, first the algo-
rithm has been run for the standard case
(a = 0.05, and b = 1.0), and next with a = 0.05,
and b = 0.0 so as to omit any influence of the pro-
ject goodness measure, vj’s. The average no. of
assignments (of the 50 runs for each case) increased
from 34 for the standard case (b = 1) to 66.5 for
b = 0.0, while the frequency of finding the optimal
solution decreased from 44 (out of 50) for b = 1 to
19 (out of 50) for b = 0.0. That is, for about 100%
extra effort the accuracy of the solution dropped
Existing Links
New Project Links
Project Links for Expansion
Fig. 2. The test network.

below 50%. This shows that vj’s have a profound
effect in guiding the search toward the optimal solu-
tion with lower effort.
Fig. 4 shows several performance measures of the
base algorithm for solving the NDP for a medium
budget level (B/C = 8330/13320 = 0.625) for three
values of a (=0.00, 0.05, and 0.20). Other parame-
ters are set at their standard levels mentioned
before. As may be seen in this figure, lower a levels
results in higher number of assignment problems to
be solved. A favorable value of a seems to be
a = 0.05, for which a good accuracy (frequency of
finding the optimal solution) exists at an acceptable
cost (average number of assignment problems
solved). Fig. 4a shows that on the average fewer
traffic assignment problems are solved as iteration
number increases, and Fig. 4b indicates that the fre-
quency of finding the optimal solution in the first
few iterations of the problem is rather high, because
of exploiting the value of the information vj. Note
that in this case it happens that only with the infor-
mation of vj the frequency of finding the optimal
solution is higher than the other two cases with
positive value of a. Fig. 4c shows that convergence
rate of the algorithm increases with a, which in turn
increases the chance of getting trapped at the local
minima. For a = 0, there is no information
exchange among ants so that the average value of
the objective function (of all solutions of each itera-
tion) alternates over iterations. And, finally, Fig. 4d
shows that the objective function of the new solu-
tions found by the algorithm in an iteration
increases as the iteration number increases, particu-
larly for the last iterations. This observation may be
used to enhance the efficiency of the algorithm. Sim-
ilar trends as discussed above for Fig. 3 are
observed for other values of budget as well.
In what follows, the effects of various improve-
ments on the base algorithms are investigated, both
separately and in combination with each other.
0
5
10
15
20
25
30
35
40
45
50
0 0.1 0.2 0.3
α
0.4 0.5 0.6
Average
no.
of
assignments
Fig. 3. Effect of a on the no. of assignment problems to be solved (B/C = 0.5, all project costs are equal).
Table 1
The performance of the base AS algorithm for different levels of B/Ca
Row B/C Budget
level
Average number of
UE assignmentsb
Frequency of finding
optimal solutionb
Value of objective
function
Bestb
Worstb
1 0.20 2700 19.2 50 76,297 76,297
2 0.32 4330 28.0 50 70,353 70,353
3 0.45 6000 25.9 50 66,650 66,650
4 0.49 6500 23.9 50 65,465 65,465
5 0.53 7075 26.6 42 64,580 65,064
6 0.63 8330 25.1 41 61,456 62,560
7 0.75 9980 20.7 50 58,839 58,839
8 0.81 10,820 18.1 39 58,829 58,839
a
a = 0.05, no. of iterations = c = 8 for all runs.

4.2. Application of hybrid ant systems
In this section the performance measures of the
proposed hybrids will be compared with the respec-
tive ones for the base algorithm. Recall that the fol-
lowing improvements have been introduced before:
Imp.1: Laying pheromone on the top 3 solutions
(AS-ACS);
Imp.2: Purposeful mutation of inferior genes
(AS-GA-TS);
Imp.3: Incorporation of Metropolis Criterion in
rejecting lower energy configurations (AS-SA);
Imp.4: Imp.2 + Imp.3;
Imp.7: Imp.1 + Imp.2 + Imp.3.
Table 2 shows the result of applying the 7 hybrid
algorithms, as well as the base algorithms, on the
network of Sioux Falls for solving a network design
problem with B/C = 0.625 (a mid-value with higher
difficulty than more extreme values), a = 0.05 (the
better value in Fig. 4), b = 1, and other parameters
as specified before. Changes in parameter values are
noted in this table. The change in the value of a
from 0.05 to 0.2 for algorithms with mutation oper-
ation, is to keep the overall effort in solving the
problem (in terms of the number of traffic assign-
ment problems solved) at a comparable level to
those without this operation so that they could be
compared better. Four measures of performance
are considered in Table 2, as in Fig. 4. These are
(a) average no. of assignments, (b) frequency of
finding the optimal solution, (c) average value of
the objective function of the solutions, and (d) aver-
age value of the objective function for the new
assignments. These information are for 50 runs of
each algorithm, and they are given for each 8 itera-
tions of the algorithms, and in total. The hybrid
Fig. 4. Effect of a on some performance measures of the base algorithm (in 50 runs, B/C = 0.625).

algorithm under the heading of the Imp.3 has been
run for the initial temperature of (3-1) c(0) = 20,000,
and (3-2) c(0) = 2. The latter version of algorithm 3
has been hybridized with the other concepts in algo-
rithms 4, 6, and 7.
As may be seen in Table 2, except for the itera-
tion 5 of the hybrids with Imp.2 that involve muta-
tion and the respective UE flow computations for
the top solutions of the previous iterations, other
hybrids experience decreasing no. of traffic assign-
ments as iteration no. increases. Fig. 5 shows the
performance of the hybrid algorithms in the space
of effort-accuracy, where effort is measured by the
no. of traffic assignments, and accuracy is measured
Table 2
Effects of various improvements upon the performance of the algorithms (50 runs)
Algorithm Base Hybrids with improvement (Imp)
Algorithm ID 0 1 2a
3 4a
5a
6 7a
Ave.
3-1b
3-2b
Measure 1: Ave. no. of traffic assignments
Iteration 1 8.6 8.7 8.6 8.7 8.9 8.8 8.7 8.4 8.7 8.7
Iteration 2 5.3 4.8 4.2 6.0 5.4 3.6 3.8 5.3 3.9 4.7
Iteration 3 3.4 3.4 1.8 3.9 3.9 1.8 1.6 3.5 2.0 2.8
Iteration 4 2.1 2.4 0.9 2.7 1.7 0.6 0.9 1.1 0.7 1.4
Iteration 5 2.2 1.9 10.0 1.3 1.3 8.3 10.0 1.0 8.4 4.9
Iteration 6 1.6 1.3 0.8 0.7 1.0 0.2 0.8 0.8 0.2 0.8
Iteration 7 1.2 1.3 0.4 0.6 0.7 0.1 0.5 0.7 0.2 0.6
Iteration 8 0.8 0.8 0.3 0.5 0.7 0.2 0.4 0.4 0.1 0.5
All iterations 25.1 24.8 27.0 24.5 23.6 23.6 26.7 21.2 24.2 24.5
Measure 2: Frequency of finding the optimal solution
Iteration 1 9 11 15 8 11 12 13 10 12 11.2
Iteration 2 12 7 9 5 6 8 10 10 12 8.8
Iteration 3 7 3 3 6 5 4 4 10 2 4.9
Iteration 4 4 3 1 9 6 2 2 4 1 3.6
Iteration 5 4 7 18 5 6 20 15 4 20 11.0
Iteration 6 4 4 0 3 6 0 1 2 0 2.2
Iteration 7 1 5 0 3 3 0 0 4 1 1.9
Iteration 8 0 2 0 2 1 0 0 0 0 0.6
All iterations 41 42 46 41 44 46 45 44 48 44.1
Measure 3: Ave. value of the objective function
Iteration 1 65,133 65,181 65,035 65,021 65,138 65,167 65,055 65,139 65,095 65,107
Iteration 2 64,903 64,965 64,447 65,204 64,925 64,700 64,387 65,070 64,595 64,800
Iteration 3 64,758 64,757 64,296 64,973 64,903 64,366 64,239 64,920 64,330 64,616
Iteration 4 64,663 64,789 64,255 64,724 64,847 64,130 64,046 64,506 64,197 64,462
Iteration 5 64,750 64,749 66,007 64,739 64,765 65,837 65,949 64,392 65,631 65,202
Iteration 6 64,693 64,677 63,970 64,730 64,696 63,913 63,873 64,441 63,851 64,316
Iteration 7 64,794 64,763 63,912 64,559 64,746 63,972 63,917 64,446 63,937 64,338
Iteration 8 64821 64,750 63,986 64,799 64,583 63,960 63,929 64,431 63,924 64,354
All iterations 64,814 64,829 64,489 64,844 64,825 64,506 64,424 64,668 64,445 64,649
Measure 4: Ave. value of the objective function for new assignments
Iteration 1 65,235 65,310 65,205 65,152 65,249 65,289 65,211 65,253 65,267 65,241
Iteration 2 65,480 65,573 65,061 65,911 65,540 65,399 65,040 65,355 65,154 65,390
Iteration 3 65,606 65,731 65,864 65,792 65,995 65,288 65,436 65,872 65,823 65,712
Iteration 4 65,810 65,727 66,523 65,213 64,788 65,026 65,489 64,769 65,285 65,403
Iteration 5 65,944 66,294 66,007 65,308 65,049 65,837 65,949 65,086 65,631 65,678
Iteration 6 66,085 65,957 66,411 65,017 64,998 66,475 66,141 65,061 64,920 65,674
Iteration 7 66,726 65,881 66,379 64,445 65,078 65,454 67,883 65,024 64,297 65,685
Iteration 8 66,254 66,160 66,954 64,821 65,324 66,465 66,944 64,929 67,211 66,118
All iterations 65,893 65,829 66,051 65,207 65,253 65,654 66,012 65,169 65,449 65,613
B/C = 8330/13,320 = 0.625, a = 0.05, b = 1.
a
a = 0.2.
b
Initial temperature, c(0), for Imp.3 has been set as (3-1) c(0) = 20,000, and (3-2) c(0) = 2.

by the frequency of finding the optimal solution (in
50 runs). According to Fig. 5, hybrid 6 (Imp.6) has
had the least average effort (21.2), and hybrid 7
(Imp.7) the highest accuracy (48). This accuracy
(48 out of 50) at a comparable effort (24.2 as com-
pared with 24.5, the average effort of all alternative
algorithms) seems a remarkable performance. This
is particularly so, if one notes that algorithm 7
reaches almost this accuracy level after iteration 5.
This phenomenon happens in algorithm 7, 5, 4,
and 2, all having Imp.2 in common. That is, the
mutation operation seems to be very effective in
identifying the optimal solution of the problem.
Other concepts seem to have had marginal effects,
when compared to the mutation of inferior genes.
It is worth noting that Ahmad (1983) has also
exploited similar mechanism and found it effective
in an integer programming context. The average
value of the objective function for the new assign-
ments in Table 2 generally rises for algorithms 2,
4, 5, and 7 (having Imp.2), but fluctuates for the
other algorithms. The average values of the objec-
tive function for the four hybrid algorithms with
Imp.2 are considerably lower than these values for
the other four algorithms without this improve-
ment, as Table 2 shows. Never-the-less, algorithm
5 (without Imp.3) shows a little lower average value
for the objective function in 50 runs than algorithm
7.
Table 3 summarizes the results of Table 2 and
compares these figures with the respective ones for
B/C ratio of 0.45 for the algorithms 0 (base), 1
(Imp.1), 2 (Imp.2), 3-2 (the better version of Imp.3
in Table 2), and 7 (the superior hybrid algorithm
in Table 2). As may be seen in Table 3, all five algo-
rithms for B/C = 0.45 happen to find the optimal
solutions in all 50 runs. In this case, algorithm 3-2
has done this with much lower effort than others.
It is worth noting that the number of feasible alter-
native networks for B/C of 0.625 and 0.45 are 781
and 398, respectively. The average number of
assignments for the five algorithms in Table 3 for
Table 3
The performance of selected algorithms for two medium budget levels
Algorithm B/C a Average number of
assignments,
all iterations
Frequency of
finding the
optimal solution
Average value of the
objective function for
the solution
0 0.45 0.05 25.9 50 66,650
1 0.45 0.05 26.6 50 66,650
2 0.45 0.2 31.6 50 66,650
3-2 0.45 0.05 20.8 50 66,650
7 0.45 0.2 29.4 50 66,650
0 0.625 0.05 25.1 41 61,532
1 0.625 0.05 24.8 42 61,525
2 0.625 0.2 27.0 46 61,477
3-2 0.625 0.05 23.7 47 61,476
7 0.625 0.2 24.2 48 61,469
0
1
2
3-1
3-2 4
5
6
7
20.0
21.0
22.0
23.0
24.0
25.0
26.0
27.0
28.0
40 41 42 43 44 45 46 47 48 49
Accuracy: freq. of finding the optimal solution
Effort:
ave.
no.
of
traffic
assignments
Fig. 5. Performance of the proposed hybrid algorithms in the space of effort-accuracy in 50 runs.

the 50 runs are 24.96 and 26.86, respectively, which
are 3.2% and 6.7% of the total feasible networks at
the respective budget levels.
5. Application of algorithms to a real case
The next case under study is the network of the
City of Mashhad, Iran, with a population of about
2.3 million in 2005. The forecasted population for
the year 2021 is about 3.0 million. The network
under study has 1298 nodes and 1726 links, and
the overall vehicle-km in morning peak hour for
the target year is about 1,450,000 (vehicles are in
pce). Total vehicle-hours spent to cross the intersec-
tions in this city is about 12000, which is around
20% of this measure in the overall network. The
problem here is to choose a subset of a given set
of intersections to be upgraded into interchanges,
so as to minimize the total user travel time for a
given budget.
Projects in this case relate to the investments in a
set of 26 intersections, which are identified based on
several criteria extracted from expert opinions,
MUTCD warrants, HCM analyses, analyses of
changes in signal settings (phases, timing, turn pro-
hibitions, etc.). The network, and the candidate pro-
jects (intersections) are shown in Fig. 6. Each
project here is a set of links and nodes representing
an interchange, which replaces another set of links
and nodes representing the current state of the inter-
section, and effectively reduce the delay at the inter-
section to zero. The construction cost of the projects
are assumed to be equal, so that budget would be
equivalent to the total number of projects intended
to be constructed. The algorithms, written in Visual
Basic, are linked to EMME/2 for the traffic assign-
ment routine. The parameters of the algorithms are
as specified before, except for a which needs recali-
bration because of the changes in the scale of the
network and demand. The followings are some of
the results obtained from the experiments done for
this case.
Table 4 compares the performance of the base
ant system algorithm with the best hybrid algorithm
(Hybrid 7). As is shown in this table, these two algo-
rithms are run for some medium budget levels of 8,
12, and 17 out of 26. The parameter a in this case is
set about 1/10 of those for the network of Sioux
Falls to suit the new order of the value of the objec-
tive function in this case. Table 4 shows that for
lower (than 0.005) values of a, Hybrid 7 performs
better than ant system and with higher accuracy in
terms of the frequency of finding the best solution
found so far, and the frequency of finding the top
10 solutions found so far.
Table 5 shows the effect of the number of candi-
date projects and the budget level on the efficiency
of the base and Hybrid 7 algorithms. In preparing
this table parameter a, and in one case the number
of iterations of the algorithm, have been adjusted
so as to obtain a comparable solution by the two
algorithms for the specified number of candidate
Roundabouts
Intersection
Fig. 6. The network of the City of Mashhad and the candidate intersection projects.

projects and budget level. The first column in Table
5 from right is the average number of network
assignments in 20 runs of the algorithm (except in
one case as mentioned in the table). These figures
may be compared with the corresponding ones in
column 2 from right in Table 5, which contains
the number of budget level permutations of the pro-
jects, in order to find the level of efforts of each algo-
rithm to find the solution. Fig. 7 shows the results of
Table 5 pictorially. As may be seen in this figure
Hybrid 7 algorithm generally outperforms ant sys-
tem algorithm in all budget levels and number of
candidate projects, yet finding better solutions
(according to Table 5). It seems that the reduction
in the number of assignments lends itself mostly to
the SA part of the Hybrid 7 algorithm than other
aspects of this algorithm. Moreover, it is worth not-
ing that as the size of the problem increases the per-
formance of Hybrid 7 algorithm increases too,
which seems to be an encouraging result for the
real-size problems.
Before passing, it might be useful to present the
results of another comparative study. In another
experiment the ant system algorithm was compared
to a genetic algorithm to solve the same problem.
The budget level has been changed from 6 to 16
Table 4
Comparison of ant system and the better hybrid algorithm at various budget levels
Row Algorithm Budget Number
of runs
a b C(0) Number
of
iterations
CPU
time
(minutes)
Average
number of
assignments
Frequency of
finding the best
solution found
Frequency of
finding top 10
solutions found
1 AS 8 2 0.005 1 – 20 23 77 2 8.5
2 AS 12 10 0.005 1 – 200 37 91.1 2 1.5
3 AS 12 10 0.001 1 – 20 71 245.3 4 3.3
4 AS 17 2 0.003 1 – 200 75 245.3 0 1
5 AS 17 2 0.001 1 – 20 83 290 1 8.5
6 H7 8 2 0.005 1 20 15 53 184.5 2 9
7 H7 12 10 0.005 1 20 20 48 165.3 10 8.1
8 H7 12 10 0.02 1 20 20 37 114 10 9
9 H7 17 2 0.003 1 20 20 61 209 1 9.5
Table 5
Effect of the number of projects and the budget level on the efficiency of the base and hybrid 7 algorithms
No. of candidate projects (N) Algorithms a Budget Possible permutations
of budget level
Average number of network assignments
10 AS 0.005 3 120 54.5
AS 0.005 5 252 60.5
AS 0.005 7 120 40.0
10 H7 0.02 3 120 52.0
H7 0.02 5 252 49.0
H7 0.02 7 120 37.0
17 AS 0.003 5 6188 62.0
AS 0.003 9 24,310 137.0
AS 0.003 12 6188 57.0
17 H7 0.02 5 6188 78.0
H7 0.02 9 24,310 97.0
H7 0.02 12 6188 52.0
26 AS 0.005 8 1,562,275 77.0
AS 0.001 12 9,657,700 873a
AS 0.001 17 3,124,550 290.0
26 H7 0.02 8 1,562,275 83.0
H7 0.02 12 9,657,700 114.5
H7 0.02 17 3,124,550 97.0
a
For 200 iterations of AS algorithm instead of 20 in other runs in the stable in order to obtain a solution comparable to those of H7.

(projects), a range suitable for the network of Mash-
had, the total number of projects was 26. Table 6
shows the results of this experiment. The projects
shown on the top of the table are those which have
been chosen by either the AS or the GA, at least
once, in the course of the experiment for various lev-
els of budget. The table also shows the solution
found by either algorithm for different budget levels,
and the respective value of the design objective func-
tion in vehicle-hour. The ‘‘1’’ in the table shows the
acceptance of the project on top of the column by
the algorithm on the left of the row, of the corre-
sponding entry of the table. As may be seen in the
table both algorithms have many projects in com-
mon for the solutions found for a specific budget
level. However, the following points are notable in
Table 6: (a) AS has consistently found better solu-
tions (with better value of the objective function)
than GA for all budget levels. (b) The projects
found by AS for various budget levels are more sta-
ble than GA which diversifies the solutions. (c) AS
solution for a budget level contained the solution
found for the previous (lower) budget level, while
GA basically diversified its solution. Moreover,
the cpu times for the GA was much more than those
of AS to reach to the solutions presented. This is the
reason for not running GA for all budget levels in
Table 6. These observations do not necessarily
imply that AS is better than GA in all circum-
stances, or that GA may not be equipped with some
tools to perform better. However, it seems that for
the nature of the problem at hand (the network
design problem) AS is a more suitable tool than
GA (which might perform better for some other
problems). It seems that flows of ants over the graph
as directed by the flow of traffic over the network
are two conformable processes, and that the latter
is not as effective when directing the GA process.
Thus, it might be more evident now that the hybrid
7 algorithm outperforms the GA-type algorithm for
the NDP.
6. Summary and conclusions
Network design problem has been, and is, an
important problem in transportation. This paper
classifies some of the existing algorithms for the
solution of this problem. Following an earlier effort
in the solution of this problem by ant system (Poor-
zahedy and Abulghasemi, 2005), this paper attempts
to improve the power of the earlier algorithm by
hybridizing it with some of the leading concepts in
search processes, such as genetic algorithm, simu-
lated annealing, and tabu search. These algorithms
are briefly described, and previous efforts in hybrid-
izing them are mentioned.
The paper, then, describes three improvements
based on the above-mentioned heuristics. Seven
hybrid algorithms have been devised based on the
base ant algorithm and these three improvement
concepts.
These seven algorithms have been applied on the
test network of Sioux Falls, South Dakota, USA. It
has been shown by various experiments on this net-
work that all three improvement concepts are effec-
tive ideas in expediting the base ant system solution
procedure.
The experiments on the network of Sioux Falls
showed the variation of some of the performance
0.0
50.0
100.0
150.0
200.0
250.0
300.0
0 2 4 6 8 10 12 14 16 18
Budget
Ave.
no.
of
assignments
in
20
runs
N=10
N=17
H7
AS
N=26
max=873 AS
AS
H7
H7
Fig. 7. Effect of the number of projects (N) and the budget level on the efficiency of ant system (AS) and hybrid 7 (H7) algorithms.

measures of the base algorithm. More specifically,
this paper has reconfirmed that:
(a) The number of traffic assignments is higher
when the ‘‘budget/total project cost’’ ratio is
around the mid-value of the total construction
cost of all projects.
(b) Higher value of a (pheromone parameter)
results lower number of assignments, which
suggests use of lower a values for low or high
B/C levels.
(c) The goodness of project j (vj) has a profound
effect upon the frequency of finding the solu-
tion of the problem, and the reduction of the
number of assignments in the design problem.
(d) Fewer traffic assignment problems are solved
as the iteration no. of the algorithm increases.
(e) The frequency of finding the optimal solution
in the first few iterations of the algorithm is
rather high.
(f) Increase of a increases the rate of convergence
of the algorithm while increasing the chance of
getting trapped at the local minima.
(g) The mutation operation (Imp.3) seems to be
very effective in identifying the optimal
solution.
(h) All hybrid algorithms have shown better per-
formance than the base Ant System.
The proposed algorithms have been applied to
solve a design problem in a real-size network of
the City of Mashhad, Iran. It has been shown that
the hybrid algorithm comprising all concepts of
AS, SA, GA, and TS perform better than the AS
alone. Never-the-less, it seems that for simpler prob-
lems (in concept, formulation, complexity, and size)
there is no need to implement a more complex algo-
rithm like Hybrid 7. The base AS would do the job.
The cost of using Hybrid 7 algorithm is justified
when more complex problems are at hand.
Acknowledgement
The authors would like to thank the Institute for
Transportation Studies and Research, at Sharif
University of Technology, Iran, for its continued
support of this study.
References
Abdulaal, M., LeBlanc, L.J., 1979. Continuous equilibrium
network design models. Transportation Research, Part B 13,
19–32.
Ahmad, N.U., 1983. An analytical decision model for resource
allocation in highway maintenance management. Transpor-
tation Research A 17A (2), 133–138.
Boyce, D.E., Farhi, A., Weischedel, R., 1973. Optimal network
design problem: A branch and bound algorithm. Environ-
ment and Planning 5, 519–533.
Cantarella, G.E., Pavone, G., Vitetta, A., Heuristics for the
network design problem. In: Ewg 2002 (the 13th Mini Euro
Conference), Bari, Italy.
Chelouah, R., Siarry, P., 2003. Genetic and Nelder-Mead
algorithms hybridized for a more accurate global
Table 6
The results of a comparison of the base AS with a GA algorithm to the case of Mashhad
Budget levela
Algorithm Project identification number Value of
objective
function
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
6 AS 1 1 1 1 1 1 69,070
7 GA 1 1 1 1 1 1 1 68,932
AS 1 1 1 1 1 1 1 68,841
8 AS 1 1 1 1 1 1 1 1 68,679
9 GA 1 1 1 1 1 1 1 1 1 68,943
AS 1 1 1 1 1 1 1 1 1 68,532
10 AS 1 1 1 1 1 1 1 1 1 1 68,344
11 GA 1 1 1 1 1 1 1 1 1 1 1 68,324
AS 1 1 1 1 1 1 1 1 1 1 1 68,202
12 GA 1 1 1 1 1 1 1 1 1 1 1 1 68,055
AS 1 1 1 1 1 1 1 1 1 1 1 1 68,039
13 GA 1 1 1 1 1 1 1 1 1 1 1 1 68,065
AS 1 1 1 1 1 1 1 1 1 1 1 1 1 67,907
14 AS 1 1 1 1 1 1 1 1 1 1 1 1 1 1 67,835
15 GA 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 67,924
AS 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 67,720
16 AS 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 67,641

optimization of continuous multiminima functions. European
Journal of Operational Research 148, 335–348.
Chen, M., Sul Alfa, A., 1991. A network design algorithm using a
stochastic incremental traﬃc assignment approach. Trans-
portation Science 25, 215–224.
Costa, D., 1995. An evolutionary tabu search algorithm and the
NHL scheduling problem. INFORM 33, 161–178.
Cree, N.D., Maher, M.J., Paechter, B., 1998. The continuous
equilibrium optimal network design problem: A genetic
approach. In: Bell, M.G.H. (Eds.), Transportation Networks:
Recent Methodological Advances. Selected Proceedings of
the 4th Euro Transportation Meeting, pp. 163–174.
Dantzig, G.D., Harvey, R.P., Lansdowne, Z.F., Robinson, D.W.,
Maier, S.F., 1979. Formulating and solving the network
design problem by decomposition. Transportation Research
13B, 5–17.
De Werra, D., Hertz, A., 1989. Tabu: A Application to Neural
Networks. OR Spectrum 11, 131–141.
Dorigo, M., Gambardella, L.M., 1997. Ant colonies for the
traveling salesman problem. BioSystems 43, 73–81.
Dorigo, M., Maniezzo, V., Colorni, A., 1991. The ant system: An
autocatalytic optimization process. Technical Report TR91-
016, Politecnico Di Milano.
Dorigo, M., Maniezzo, V., Colorni, A., 1996. The ant system:
Optimization by a colony of cooperating agent. IEEE
Transactions on Systems, Man and Cybernetics, Part B 26
(1), 1–13.
Fleurent, C., Ferland, J., 1999. Genetic Hybrids for the quadratic
assignment problem. DIMACS Series in Mathematics Theo-
retical Computer Science 16, 190–206.
Friesz, T.L., Cho, H.J., Mehta, N.J., Tobin, R.L., Anadalingam,
G., 1992. A simulated annealing approach to the network
design problem with variational inequality constraints. Trans-
portation Science 26 (1), 18–26.
Glover, F., Kelly, J.P., Laguna, M., 1995. Genetic algorithms and
tabu search: Hybrids for optimization. Computer Opera-
tions Research 22 (1), 111–134.
Goldberg, D.E., 1989. Genetic Algorithm Optimization.
Addison-Wesley Publishing Company, Reading, Mass.
Greistorfer, P., 1998. Hybrid genetic tabu search for cyclic
scheduling problem. In: Meta-Heuristics Advances and
Trends in Local Search Paradigms for Optimization. Klower,
Boston.
Haghani, A.E., Daskin, M.S., 1983. Network design application
of an extraction algorithm for network aggregation. Trans-
portation Research Record 944, 37–46.
Hertz, A., Widmer, M., 2003. Guidelines for the use of
metaheuristics in combinatorial optimization. European
Journal of Operational Research 151, 247–252.
Hoang, H.H., 1982. Topological optimization of networks: A
nonlinear mixed integer model employing generalized benders
decomposition. IEEE Transaction on Automatic Control 27,
164–169.
Holmberg, K., Hellstrand, J., 1998. Solving the uncapacitated
network design problem by a Lagrangian heuristic and
branch-and-bound. Operations Research 46 (2), 247–259.
Kim, H., Nara, K., Gen, M., 1994. A method for maintenance
scheduling using GA combined with SA. Computers Indus-
trial Engineering 27, 477–486.
LeBlanc, L.J., 1975. An algorithm for discrete network design
problem. Transportation Science 9, 183–199.
Lee, C.K., Yang, K.I., 1994. Network design of one-way streets
with simulated annealing. Papers in Regional Science 32 (2),
119–134.
Magnanti, T.L., Wong, R.T., 1984. Network design and trans-
portation planning: models and algorithms. Transportation
Science 18 (1), 1–55.
Maniezzo, V., Gambardella, L.M., Deluigi, F., 2003. Ant colony
optimization. In: Onwubolu, G.C., Babu, B.V. (Eds.), New
Optimization Techniques in Engineering. Springer-Verlag
(www.idsia.chnlucanACO.2004.pdf).
Onwubolu, G.C., 2002. Emerging Optimization Techniques in
Production Planning and Control. Imperial College Press,
London.
Poorzahedy, H., Abulghasemi, F., 2005. Application of ant
system to network design problem. Transportation 32, 251–
273.
Poorzahedy, H., Turnquist, M.A., 1982. Approximate Algo-
rithms for the Discrete Network Design Problem. Transpor-
tation Research, Part B 16, 45–56.
Resende, M.G.C., Pardalos, P.M., Eksioglu, S.D., 1999. Parallel
Metaheuristics for Combinatorial Optimization, www.
research.att.com/mgcr/doc/parheur.pdf.
Sadek, A.W., 2001. A hybrid simulated annealing and case-based
reasoning approach for computationally intensive transpor-
tation problem: Rationale and design issues. In: TRB, 80th
Annual Meeting, Washington, DC.
Salmon, P., Sibani, P., Frost, R., 2002. Facts, Conjectures and
Improvements for Simulated Annealing. SIAM Society of
Industrial Applied Mathematics, Philadelphia, PA.
Solanki, R.S., Gorti, J.K., Southworth, F., 1998. Using decom-
position in large-scale highway network design with quasi-
optimization heuristic. Transportation Research, Part B 32,
127–140.
Steenbrink, P.A., 1974a. Optimization of Transport Network.
John Wiley, NewYork.
Steenbrink, P.A., 1974b. Transportation network optimization in
the Dutch integral transportation study. Transportation
Research 8, 11–27.
Taillard, E.D., Gambardella, L.M., 1997. Adaptive memories for
the quadratic assignment problem. Technical Report IDSIA-
79-97, Lugano, Switzerland.
Wong, R.T., 1984. Introduction and Recent Advances in
Network Design Models and Algorithms. In: Florian, M.
(Ed.), Transportation Planning Models. North-Holland,
Amsterdam.
Yang, H., Bell, M.G.H., 1998. Models and algorithms for road
network design: A review and some new developments.
Transport Review 18 (3), 257–278.
Yin, Y., 2000. Genetic algorithm-based approach for bilevel
programming models. Journal of Transportation Engineer-
ing: ASCE 26 (2), 115–120.

Hybrid Meta-Heuristic Algorithms For Solving Network Design Problem

Recommended

Recommended

More Related Content

Similar to Hybrid Meta-Heuristic Algorithms For Solving Network Design Problem

Similar to Hybrid Meta-Heuristic Algorithms For Solving Network Design Problem (20)

More from Alana Cartwright

More from Alana Cartwright (20)

Recently uploaded

Recently uploaded (20)

Hybrid Meta-Heuristic Algorithms For Solving Network Design Problem